What if 538 is right, and Romney loses by 60+ electoral votes

Ah, I see - you’re saying it doesn’t go back far enough.

Perhaps. But maybe elections are just too different before 1980. The fact that this model predicts all of them since then isn’t too shabby. And maybe it predicts more correctly before 1980, just not all of them.

I don’t know what a good standard is.

I didn’t say I believe the model. I just said going back into the past to test it is a legitimate way to test it.

I’m not stupid. My money is on Obama.

I think the issue is that it hasn’t really predicted an election that wasn’t also a part of the data set on which it is created. Suppose it’s wrong this time. Then the people who created it tweak the variables such that it would have been correct, then they say that their new model is perfect from 1980 through 2012. Would you trust it then? It still won’t have predicted anything.

As an incidental note, fivethirtyeight is only proven “right” with a statement like “Obama has a 72.4% chance of getting 330 or more electoral votes” if we run the election several hundred times and observe this result with an acceptable confidence level.

[/stats nerd]

Of course not. None of them are. They make models by looking at past data, after all, and seeing which correlations are the strongest.

Of course I would. The model got better. Before it predicted all but one election since 1980, and this time it would have predicted all of them. It’s how models work. Not perfect, but good.

But your’e right - there’s no point in a model that fails in the future alot, even if it works in the past. That means something is wrong with it, or something has changed since it was created.

True… I’ve heard this argument taken to an extreme: one stats classmate of mine insisted that you can’t talk meaningfully about the “probabilities” or “odds” of winning or losing at roulette, since each game is unique and has no relationship to any other.

The standard response is that each spin in roulette is so very similar to every other that you can treat them as a population of “the same” game. It isn’t absolutely true; each roulette wheel will have some very tiny physical bias built in.

The key rebuttal is that stats actually do work in the real world – roulette casinos don’t tend to lose money!

Looking at what time someone shows up for work is usually a very good predictor of when they will show up for work tomorrow.

I think they meant that you can’t predict the outcome of a game by the previous games. Which is true, statistically (assuming there’s no bias in the wheel).

Casinos don’t lose money on roulette because they have a zero (and in the U.S. a double zero) that only the house wins.

Every politician’s dream. And why Sheldon Adelson has so much money to spend.

Now that you mention it, shouldn’t the probabilities swing hard one way or another and spend most of their time around 99%? It would only be somewhere between 10-90% for a very close election, right? Or am I missing something?

I don`t think there’s anything particularly terrible about thinking this is true (except that you’re wrong). It only becomes embarrassing if you publish a model based on it and attract national attention all on the hopes that not enough Americans will understand statistics well enough to realize why your model is bs.

I don’t fully understand your question, but if a coin is slightly biased, we could say it has a 70% chance of landing on “heads”. Fivethirtyeight is calling this election (as best it can predict at this moment) as biased in Obama’s favour.

As the date approaches, I can picture the 70% becoming 90%, if events continue to run in Obama’s favour and the sizable “undecided” cohort starts to side with him, but there’s enough randomness in an election that this might not happen.

This really depends on what you want to use the model for. If you really are concerned with accurate prediction, then you have to make out-of-sample predictions. It’s not a great idea to test the model on the same set you trained it on because it’s very easy to overfit a predictive model. This is especially true when the number of independent variables approaches the number of observations.

So explain why it’s B.S.

It’s perfectly legitimate to look at past correlations and try to find causation and then test your hypothesis in the future. In fact, that’s the scientific method right there. Doesn’t mean the model is right, but the method is sound.

Moderate Republican. of course, more of the Rockefeller variety. He can work with the other side, he accepts science, and everything I’ve read about him makes me think that he’d respond to the facts, not ideology. I’m sure he’d raise taxes. There were lots of Republican presidents who didn’t drive the country into the crapper, back in the hazy past.

Heck, give Bush senior another term.

Spelling Obama with a zero means I should take everything you say extra seriously.

How’s Free Republic these days?

What’s BS is claiming the model is a strong one solely based on its being able to account for the very data it is based on.

Say that every Presidential election since 1980 had been won by the guy with the most vowels in his name. You figured that out by going back and counting vowels and seeing the correlation. Then you say, “Well, Obama is going to win in 2012 because he has more vowels than Romney and this model would have predicted the right outcome since 1980.”

Do you see why that doesn’t work? It’s not a prediction if you simply tailor it to the results you know happened. You have to test it before you can say that it has any ability to predict anything and you can’t test it by using the same data that you used to build it.

Or a different example: You talk to a friend and you hear what the friend has to say about the election. Then you repeat what they said back to them and they agree. Do you see why you couldn’t then say that all of the people you asked agreed with what you said?

No black incumbent has ever won an election. This model has 100% accuracy going back way past 1980, can I get mentioned in the press now?